Experts lie — and the worst part is that they almost never are

In Freakonomics, Steven Levitt and Stephen Dubner showed something simple and uncomfortable: when a real estate agent sells their own home, they leave it on the market about ten days longer and close for about 3% more than when selling yours. They are not dishonest. They simply know something you do not, and use it in line with their interest — which is not identical to yours. That sentence explains half a century of errors that cost lives: the expert rarely lies; they respond to incentives you cannot see.

1. Philosophy got there first — by several different roads

The intuition that the expert misleads was not born in behavioural economics. It has a long lineage, and each link adds a mechanism.

Socrates, in the Apology, reports seeking out craftsmen expecting to find wisdom. He found real competence — and, alongside it, the flaw that struck him: because they mastered their own trade well, they believed themselves knowledgeable about the greatest questions. It is the first description of the expert who overruns the boundary of their own competence, which is the commonest error of all.

Francis Bacon, in the Novum Organum (1620), catalogued the idola theatri — the idols of the theatre: received philosophical systems accepted as dogma because they come from authority rather than because they were verified. Bacon was describing the cognitive cost of trusting the school instead of the experiment.

Adam Smith, in 1776, was harsher and more modern: “People of the same trade seldom meet together, even for merriment and diversion, but the conversation ends in a conspiracy against the public, or in some contrivance to raise prices.” Note that Smith is not talking about villains — he is describing an outcome that emerges whenever identical interests gather.

George Bernard Shaw compressed it all into one line in The Doctor’s Dilemma (1906): “All professions are conspiracies against the laity.” The play is about doctors choosing who will live, and about how professional prestige protects itself.

In the twentieth century the argument acquired technical apparatus. Thomas Kuhn (1962) showed that the normal scientific community does not try to refute its own paradigm — it solves puzzles inside it, and resists anomalies until it no longer can. George Stigler, a future Nobel laureate, formulated regulatory capture in 1971: as a rule, he argued, regulation is acquired by the industry and operated for its benefit. Ivan Illich, in Medical Nemesis (1975), coined the idea of disabling professions — those that create the dependency they then offer to treat.

And there is empirical verification. Philip Tetlock tracked, over twenty years, the forecasts of 284 political and economic experts — more than 80,000 predictions: on average, barely better than chance. Tetlock himself rejects the popular “dart-throwing chimpanzee” summary, and the caveat matters — the experts beat chance over short horizons, and only approached the chimpanzee when projecting three to five years out. The precise lesson is not that experts know nothing; it is that their confidence does not fall when their ability to be right falls.

Finally, John Ioannidis, in 2005, published in PLoS Medicine the most cited article in the journal’s history, with a title that became a proverb: Why Most Published Research Findings Are False. The argument is not rhetorical but statistical — given typical sample sizes, the number of hypotheses tested and analytical flexibility, in many fields a published claim is more likely to be false than true.

2. Tobacco: the case that became the manual

If there is one example where “lying” is the right word, this is it. In January 1954 the cigarette industry published the Frank Statement to Cigarette Smokers in hundreds of American newspapers, promising to fund independent health research, and created the Tobacco Industry Research Committee for the purpose. What the committee produced over the following decades was not truth: it was controversy.

The strategy is written down, without euphemism, in an internal Brown & Williamson memo from 1969:

Notice what is being said. The stated goal was not to prove that cigarettes are good for you. It was to prevent consensus from forming — because for a seller, doubt works as well as approval. By 1964 the Surgeon General’s report had already concluded, across more than seven thousand papers, that smoking causes lung cancer. The manufacture of doubt lasted, even so, some 45 years, and was only dismantled when internal documents surfaced in litigation.

Keep this case in mind: it is the mould. All the others repeat parts of it.

3. Sugar: the day fat took the blame

In 2016, Cristin Kearns and colleagues published in JAMA Internal Medicine an analysis of internal sugar industry documents. The finding: in 1965 the Sugar Research Foundation commissioned an internal review called Project 226 from researchers at the Harvard School of Public Health — among them Mark Hegsted and Robert McGandy — paying the equivalent of about 48,000 dollars in 2016 money. The commission had a declared target: to neutralise the studies linking sucrose to coronary disease.

The review appeared in 1967, in the New England Journal of Medicine. It disqualified the work incriminating sugar and granted merit only to work pointing at fat and cholesterol. The funding was not disclosed — at the time the journal did not require it. Mark Hegsted would go on to hold an influential position in shaping American dietary guidelines.

What came next is the answer to the complaint about nutrition — that one day they recommend a food and the next they condemn it. It is not scientific fickleness: it is the trail of a badly framed question with money inside it.

• Fat becomes the villain and industry answers with “low fat” products, in which the fat removed is compensated with sugar.
• Margarine is recommended over butter. Decades later it emerges that the trans fat in partially hydrogenated oils is worse than what it replaced — the US revoked its safe status in 2015, with a ban following.
• The egg is condemned for cholesterol and later absolved: the 300 mg daily limit was dropped from the American guidelines in 2015.

Three reversals, one common origin: strong conclusions drawn from weak evidence, defended by authority, with a commercial interest anchored to each of them.

4. The pharmaceutical industry: when the error has bodies

Here the pattern gets more expensive, because the product goes into the vein.

Thalidomide (1957–1961)

Sold as a safe sedative for nausea in pregnancy, it produced thousands of children with severe malformations before being withdrawn. It is the founding milestone of modern pharmacovigilance — and the proof that “no evidence of harm” is not the same as “evidence of no harm”.

Vioxx (1999–2004)

Merck’s anti-inflammatory was withdrawn worldwide on 30 September 2004, after it was confirmed that it doubled the risk of heart attack and stroke with prolonged use. The VIGOR study, in 2000, already showed excess cardiovascular events, interpreted benignly at the time. David Graham, a researcher at the FDA itself, estimated that about 88,000 Americans had heart attacks because of the drug, of whom around 38,000 died. It was not one isolated expert who failed: the manufacturer, peer review and the regulator failed at the same time.

OxyContin and the opioid epidemic

This is the most instructive example of all, because it shows how a sentence becomes science. In 1980, Jane Porter and Hershel Jick sent the New England Journal of Medicine a one-paragraph letter noting a low rate of dependence among hospitalised, monitored patients receiving opioids. It was not a study. It was a letter.

That letter was cited hundreds of times as if it were evidence that prescribed opioids rarely addict, and it became the basis of marketing: OxyContin advertising claimed the rate of addiction among patients treated by doctors was “much less than 1%”. In 2007, Purdue Pharma and three of its executives pleaded guilty to misleading regulators, doctors and patients about the drug’s addictive potential. Between the letter and the confession, hundreds of thousands of people died.

And the silent mechanism: the study that never appears

None of these cases requires an evil scientist. All it takes is publication bias: whoever funds the work is not obliged to publish what they disliked. The effect is arithmetical — if half the studies come out negative and almost only the positive ones are published, the entire literature lies without a single line being false. That is why prospective trial registration and the requirement to publish negative results were the most important reforms in medicine of the last two decades.

5. Alcohol: the safe dose that never existed

For thirty years, the official answer to “how much can I drink?” was a curve. The J-curve showed that people who drank a little lived longer than people who drank nothing — and the daily glass of red wine entered popular culture as medical advice.

The problem was methodological, and it has a name: abstainer bias, or the “sick quitter” effect. The comparison group — those who do not drink — was contaminated by people who stopped drinking because they were already ill. Comparing moderate drinkers with that group means comparing healthy people with sick people, and calling the difference a benefit of alcohol.

The meta-analysis by Tim Stockwell and colleagues examined 107 studies, with more than 4.8 million participants: only 21 were free of some form of that bias. Once the defect is corrected, the longevity advantage of the moderate drinker disappears. Mendelian randomisation studies, which use genetic variation and escape this kind of confounding, likewise found no cardiovascular protection.

In January 2023, the World Health Organization published the blunt conclusion: there is no level of alcohol consumption that is safe for health. Ethanol has been classified by the IARC as a Group 1 carcinogen — the asbestos and tobacco category — since 1988. Cancer risk begins to rise from the first sip; what exists is lower risk, not zero risk.

And the link to the rest of this article has a date and a price tag. MACH15 was to be the great randomised trial that would finally answer whether moderate drinking protects the heart. A budget of roughly 100 million dollars, run under the umbrella of the American NIH — with approximately two thirds of the funding coming from five industry giants: Anheuser-Busch InBev, Carlsberg, Diageo, Heineken and Pernod Ricard. In June 2018, the NIH cancelled the study after an internal investigation concluded that staff had solicited the industry money and that the trial design was tilted — a primary endpoint favourable to alcohol and insufficient attention to non-cardiovascular risks such as cancer.

In other words: the study that would answer the question was designed, in part, by those who had a preferred answer. This is the same film as 1954, with a different product.

6. The pandemic: the cost of asserting more than you know

This is the most recent example and the most keenly felt — and for that reason the one that demands the most discipline, because here the easy error is to commit the same sin in reverse.

Start with what is solid: the vaccines sharply reduced severe disease and death. That showed up in the randomised trials and then in population data, and denying it would be exactly what this article criticises — asserting more than the evidence supports. That was not the problem.

The problem was transmission, and the certainty with which it was discussed.

On 29 March 2021, the director of the American CDC, Rochelle Walensky, said on national television that the data suggested vaccinated people “do not carry the virus, do not get sick”. Three days later, on 1 April, the CDC itself walked it back through the press: the director had spoken “broadly” and the evidence on transmission “was not clear”. The retraction reached a fraction of the audience of the claim.

In October 2022, at the European Parliament, the executive Janine Small confirmed that the Pfizer vaccine had not been tested for transmission before launch. Fairness requires saying that Pfizer had never claimed otherwise — the trials measured symptomatic disease, and that was published. And that is where the serious point lies: the promise that circulated did not come from the label, it came from public communication, and nobody in authority corrected it while it was useful.

In July 2021, the Provincetown outbreak in Massachusetts, published by the CDC itself, made correction unavoidable: of 469 cases, 346 (74%) were in fully vaccinated people, and measured viral load was equivalent between vaccinated and unvaccinated. The Delta variant had changed the picture; the public message took longer to change with it.

The rare harms are real and were acknowledged. The Nordic study published in JAMA Cardiology in 2022, covering 23 million residents, confirmed increased risk of myocarditis and pericarditis after mRNA vaccines, concentrated in young males after the second dose: between 4 and 7 excess events per 100,000 vaccinated with the Pfizer product, and between 9 and 28 per 100,000 with Moderna’s. These are small numbers — and they are true numbers, which exist and were measured. Several Nordic countries restricted Moderna in young people because of them.

For viral vector vaccines, AstraZeneca acknowledged in UK court documents in April 2024 that Vaxzevria can cause thrombosis with thrombocytopenia syndrome; in May 2024 the company withdrew the product worldwide, citing commercial reasons and falling demand.

Two caveats of honesty, because without them the argument loses its value. First: Denmark did not ban the vaccine for people under 50 — it stopped actively inviting that group in booster campaigns, concentrating the effort on older and higher-risk people, and anyone who wanted the vaccine could still have it. The distorted version circulates widely; using it would repeat the very error we are denouncing. Second: a rare risk does not cancel the benefit for someone at high risk of dying — what it makes untenable is the zero-risk discourse, and it was that discourse that was used to make vaccination compulsory or a condition of access to venues, school and employment.

Because this is the genuinely uncomfortable conclusion: the coercion was built on the part of the promise that had no backing. Requiring proof of vaccination to enter a restaurant only makes sense if the vaccinated person does not transmit. When it became known that they did, the requirement stayed in place for a while, and debate about it kept being treated as denialism. The experts’ error here was not the data — it was the certainty, and the closing of the space where legitimate doubt could have been voiced.

7. The pattern: five mechanisms, none of them needing a conspiracy

Put tobacco, sugar, Vioxx, opioids, alcohol and the pandemic together and what repeats is not villains. It is machinery:

• Whoever pays chooses the question. No need to falsify results: it is enough to fund the questions whose answers suit you and not fund the others. Project 226 and MACH15 are the same move, 50 years apart.
• What comes out negative never appears. Publication bias makes the entire literature lie without any single paper being false.
• The regulator is captured. Stigler described it; Vioxx demonstrated it. The body meant to watch lives alongside, hires from and is funded by those it watches.
• Changing your mind in public is expensive. Careers are built on theses. Admitting error after twenty years of defending one is not only an intellectual cost but a professional one — and the incentive pushes towards restating.
• Certainty is rewarded; doubt is punished. Television invitations, funding and prestige go to those who speak firmly. Tetlock measured the result: expert confidence does not fall when the ability to be right falls.

Add the five and you get a system that produces error with a direction — always favouring whoever funds — without anyone needing to sit in a room and agree on anything. It is worse than conspiracy, because a conspiracy can be undone by arresting the guilty. This can only be undone by changing incentives.

8. The wrong conclusion — and why it is dangerous

The obvious exit from this article would be: trust no one. That would be the wrong reading, and it is worth saying why.

It was scientists who uncovered the tobacco fraud. It was a researcher reading internal documents and publishing in a medical journal who revealed the sugar industry’s payment to Harvard. It was epidemiologists reanalysing 107 studies who brought down the alcohol J-curve. It was a Stanford professor of medicine, publishing in a scientific journal, who showed that most published findings are false.

In every case, the corrective for bad science was more science, never less. Anyone who concludes “therefore there is no truth” hands the field precisely to those who profit from doubt — which is, literally, the product Brown & Williamson said it was selling. The scepticism that serves is the kind that demands method, not the kind that abandons the idea of evidence.

What the record recommends is something else, more modest and more demanding: stop delegating blindly. Not to replace the expert — to audit them.

9. What changed: auditing the expert got cheap

Until recently, that advice was empty. “Read the original study” is easy to say to someone without journal access, without technical English, who does not know what a confidence interval is or where to find the conflict-of-interest statement. Auditing an expert was itself expert work — and experts charge by the hour.

That cost is exactly what artificial intelligence has collapsed. Not because it knows more than the best specialists — on average it knows less, and about the specific case of your body or your lawsuit, far less. But because it performs, in minutes and without fees, the tasks that used to separate the layperson from the primary source:

• Read the study, not the headline. The news says “coffee reduces mortality”; the study says observational cohort, weak association, no control for smoking. The distance between those two sentences is where nearly all the deception lives.
• Ask who paid. Practically every serious paper today carries a funding and conflict-of-interest statement. Almost nobody reads that section. Now it costs one question.
• Check the outcome measured. Did the study measure death and heart attack, or a surrogate marker — cholesterol, bone density, viral load? Plenty of approved drugs improve the number without improving the life.
• Ask for the systematic review, not the isolated study. One study is a large anecdote. What matters is the body of work, and whether it is consistent.
• Hear the opposing case at its strongest. Explicitly ask for the best argument against what you have just concluded. That is the thing an expert paid to defend a position will never give you for free.
• Consult many without paying each. A second opinion used to cost a consultation; a third, another. Today you can confront the reasoning of several schools before choosing whom to pay — and arrive at the appointment knowing what to ask.

10. And who audits the AI?

It would be incoherent to end an article about overconfidence by asking for blind confidence in something else. Artificial intelligence has exactly the defects this text has described, in a new form:

• It learned from the biased literature. If an entire decade published that fat was the villain, that is what it absorbed. It reproduces consensus — including when the consensus was bought.
• It is trained to please. Agreeing with whoever asks is a measurable bias of these models. If you frame the question with the answer already inside it, there is a good chance you will get your own opinion back, better written.
• It invents sources. A study that does not exist, a precedent never decided, a citation with perfectly plausible author and year. The error is rare enough not to show up in casual use and frequent enough to destroy a piece of work.

What corrects all three is the same method, applied to it: ask for the source and open the source. Ask for the counter-argument before asking for the supporting one. Do not ask “why is X bad for you?”, ask “what is known about X, and where is the evidence weak?”. And check that what it cited exists — because the tool that collapses the cost of checking also collapses the cost of inventing.

Notice that none of this is new. It is the method Bacon proposed against the idols of the theatre: do not accept the authority of the school, go to the experiment. What changed in 2026 was not the method. It was the price.